Industries / Transport and vehicles / Automotive
Software in the car
This work writes and approves the code and electronics inside a vehicle: engine computers, in-car networks, software updates, driver assistance and protection against hacking.
7
Milestones achieved, from 1977 to 2024.
2
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1977
A microprocessor times the spark
GM's MISAR system used a digital microprocessor to set ignition timing and was standard equipment on the 1977 Oldsmobile Toronado. The SAE technical paper describing it calls the Toronado the first production car equipped with an electronic digital microprocessor.
Credit
Delco-Remy Division of General Motors; Thomas W. Evernham and Donald G. Guetersloh, the paper's authors
1986
One network for many computers
In February 1986 Bosch introduced the Controller Area Network (CAN), a shared data line over which a car's control units exchange messages. Mercedes-Benz has used CAN in its upper-class cars since 1991, and ISO published it as the standard ISO 11898 in November 1993, by the account of the CAN in Automation users' group.
Credit
Robert Bosch GmbH
2003
Shared software architecture
Car makers and suppliers agreed on AUTOSAR, an open standard for how the software inside control units is structured. The core partners signed the first development agreement in July 2003; Ford, PSA and Toyota joined later that year and General Motors in 2004.
Credit
BMW, Bosch, Continental, DaimlerChrysler and Volkswagen, with Siemens VDO
2015
Jeep hacked from afar
Researchers Charlie Miller and Chris Valasek took remote control of some functions of an unmodified 2014 Jeep Cherokee through its Uconnect radio. Fiat Chrysler recalled about 1.4 million vehicles, and the fix went out on USB drives mailed to owners. A year earlier Tesla had fixed a recall with an over-the-air (OTA) update, software sent to the car over the mobile network.
LessonA connected car needs a secure way to be fixed remotely.
Credit
Miller and Valasek; NHTSA; Fiat Chrysler
2020
Rules for cyber and updates
In June 2020 the UN body that sets vehicle rules, UNECE's WP.29, adopted Regulation 155 on cybersecurity and Regulation 156 on software updates, under which a car maker must show before sale that it manages both. In the EU, protection against cyberattacks became required for new vehicle types from 6 July 2022 and for all newly registered vehicles from 7 July 2024, and the software-update rules reach all new vehicles from 7 July 2026.
Credit
UNECE WP.29; the European Union
2021
First automated-driving approvals
Japan's transport ministry approved Honda's Traffic Jam Pilot, a Level 3 system, meaning the car drives itself only in set conditions such as slow highway traffic, and Honda began leasing 100 Legend sedans with it on 5 March 2021. In December 2021 Germany's vehicle authority, the KBA, approved Mercedes-Benz Drive Pilot under UN Regulation 157 for use at up to 60 km/h.
LessonThe first approvals came with narrow limits on speed and road type.
Credit
Honda and Japan's Ministry of Land, Infrastructure, Transport and Tourism; Mercedes-Benz and the KBA
2024
In-house software falls short
Volkswagen's software company, Cariad, was building a common platform for the group's brands, and Porsche said in 2022 that delays in its E3 1.2 platform held back the electric Macan. In November 2024 Volkswagen formed a joint venture with Rivian, worth up to $5.8 billion, to use Rivian's zonal design, which groups computers and wiring by area of the car.
LessonA large car maker paid a younger rival for an electronics base it had struggled to build itself.
Credit
Volkswagen Group, Cariad, Porsche and Rivian
Watch for
EU-wide approval for Tesla FSD
On 10 April 2026 the Dutch vehicle authority RDW approved Tesla's driver-assistance system FSD (Supervised), valid for now only in the Netherlands. RDW states that a car with it is not self-driving, and it has applied to the European Commission for approval across the EU; Electrive reported in September 2026 that seven EU countries had adopted the Dutch approval and that an EU vote was not expected before December.
Achieved whenThe EU's Technical Committee on Motor Vehicles approves it for the whole EU.
Claimed by
Tesla, through RDW's application, April 2026
China's driver-assistance standard
China issued GB 47955-2026, a mandatory national standard with safety rules for combined driver assistance, meaning systems that steer and control speed together while the driver supervises. It takes effect on 1 January 2027.
Achieved whenThe standard is in force and new Chinese type approvals are tested against it.
Claimed by
China's Ministry of Industry and Information Technology, July 2026
Open
01
Machine-checked openpilot safety rules
Computer-checked proofs that the open-source safety code in comma.ai's opendbc project always enforces its own rules. That code sits between the openpilot driver-assistance software and the car. Example rules: the device stays silent on the car's network until a safety mode is chosen, the brake pedal always cancels control, and each car's limits hold. The work covers source code and simulation only, and no modified firmware goes into a car on public roads.
The briefCredit · unclaimed02
Plain-language software recall ledger
An open, versioned table of every US vehicle recall since 2020, with labels NHTSA does not store as fields: whether software caused it, whether the fix is an over-the-air update, a dealer reflash or a part swap, and whether it replaces an earlier recall. Each recall also gets a two-sentence summary in plain English.
The briefCredit · unclaimed03
Rate hazards like an engineer
A model or agent that reads a hazard case and assigns its ASIL, the risk level from A (lowest) to D (highest) in the ISO 26262 functional-safety standard, scored on the public SAFARI benchmark. The step being tested, hazard analysis and risk assessment (HARA), lists what can go wrong with a vehicle function and rates each hazard.
The briefCredit · unclaimed
Open 01 · the brief
Machine-checked openpilot safety rules
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Computer-checked proofs that the open-source safety code in comma.ai's opendbc project always enforces its own rules. That code sits between the openpilot driver-assistance software and the car. Example rules: the device stays silent on the car's network until a safety mode is chosen, the brake pedal always cancels control, and each car's limits hold. The work covers source code and simulation only, and no modified firmware goes into a car on public roads.
Why it is still open
The project checks its safety code with unit tests for each supported car, MISRA C rule checks (a coding rulebook for safety-critical software), mutation tests and 100% line coverage, and its README does not mention formal proof. Searches for formal verification of this code found none.
Built with what exists
The opendbc repository; CBMC or Frama-C, open-source tools that prove properties of C code; an AI coding agent to draft proof harnesses for a person to review (see the Notes "Proven in four weeks" and "Findings that cite notes").
First step
Clone opendbc, build its safety tests, and write one CBMC harness proving that the silent mode lets no message out.
Done looks like
A public repository where one command runs proofs for at least three safety modes, each proving that the device is silent by default, that the brake cancels control and that the car's limits hold, with any counterexample filed with the project.
Open 02 · the brief
Plain-language software recall ledger
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open, versioned table of every US vehicle recall since 2020, with labels NHTSA does not store as fields: whether software caused it, whether the fix is an over-the-air update, a dealer reflash or a part swap, and whether it replaces an earlier recall. Each recall also gets a two-sentence summary in plain English.
Why it is still open
NHTSA's public recall file stores the fix only as a free-text "corrective summary", with no field for the type of remedy, and the independent trackers found offer no downloadable labeled dataset.
Built with what exists
NHTSA's recall flat file and recall reports; a small model fine-tuned on one graphics chip (see the Note "One GPU to fine-tune") or run locally (see the Note "A first stable number for local AI"); a second pass that flags doubtful labels (see the Note "Seven more flagged").
First step
Download NHTSA's recall file, filter it to 2025, and hand-label 100 recalls.
Done looks like
Labels for all recalls from 2020 to 2025, with two people hand-labeling a random sample of 300 and the model's remedy-type labels matching theirs at least 95% of the time; data and code published under an open license.
Open 03 · the brief
Rate hazards like an engineer
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A model or agent that reads a hazard case and assigns its ASIL, the risk level from A (lowest) to D (highest) in the ISO 26262 functional-safety standard, scored on the public SAFARI benchmark. The step being tested, hazard analysis and risk assessment (HARA), lists what can go wrong with a vehicle function and rates each hazard.
Why it is still open
SAFARI's authors tested nine leading AI models on 3,000 anonymized industrial HARA cases, and the best reached an ASIL macro-F1, an accuracy score from 0 to 1, of only 0.261.
Built with what exists
The SAFARI dataset on GitHub; a small model fine-tuned on one graphics chip (see the Note "One GPU to fine-tune"); a reviewer pass that flags doubtful ratings for a person to check (see the Note "Seven more flagged").
First step
Download SAFARI, reproduce one baseline score from the paper, and list which ASIL levels the baseline confuses.
Done looks like
An ASIL macro-F1 of at least 0.5 on SAFARI, scored with the paper's own method, with code and predictions published so another group can reproduce the number.